Stadium crowd counting method based on peer to peer network
Xingya Yan, Yujiao Ding · 2023
The crowd base of the stadium can directly reflect the degree of congestion of the field, and counting, analyzing and evaluating the crowd through Internet artificial intelligence is helpful to improve the management performance of the stadium. The traditional stadium crowd counting is usually conducted by the administrator directly observing the number of athletes, or after the monitoring video rough counting, this kind of crowd counting is time-consuming and laborious, the accuracy is low, and it is hard to real-time accurately reflect the crowd density in the stadium and the management problems existing in the stadium in real time. Therefore, this thesis puts forward a stadium people counting method based on point-to-point network. First, Mosaic technology was used to improve the robustness of the subsequent detection model, and then yolov5 algorithm was used to obtain human head detection points, and improvements were made based on the point-to-point network model to increase the accuracy of people counting. VGG-19 network model was used to obtain the final count results for the key points of the head obtained through the point-to-point network. The experimental results on ShanghaiTech, UCF-QNRF_ECCV18, jhu_crowd_v2.0 and NWPU data sets show that compared with the previous recognition and counting models, the average accuracy of counting has improved, and MAE reaches 55.74%, which can be effectively in a position to people counting in various scenes in the stadium.